To test the study hypotheses analytic statistics were conducted using IBM SPSS Statistics 24, in the form of multiple linear regression. The
assumptions for this model were met within the data set as described above. This particular statistical method was chosen because there were multiple predictor variables, and multiple linear regression allows for the
determination of which predictors are significantly associated with the
independent (criterion) variable, while taking into account that the predictors may be related to each other (Howitt & Cramer, 2005). Within this approach the method of hierarchical linear regression was employed due to there being specific hypotheses developed based on theories derived from previous research in relation to the role of tumour localisation in phonemic fluency performance (Field, 2009). In line with this the localisation variable was added to the model in the first stage of the regression, with the predictor variables whose assumed influence was more clearly supported by prior
research entered second (gender and education); additional predictor variables added in the second stage. Significance was assessed at an α level of 0.05.
Due to the unexpected loss of usable data through the data cleaning and categorising process, particularly when considering the hierarchical multiple regression analyses (reducing from 123 to 80 cases) with a relatively high number of variables, a post hoc power analysis is
recommended to determine the chance of the results being due to a type 2 error. G*Power (Faul et al., 2009) was used to calculate a post hoc analysis for each of the regression analyses based on their number of predictors, their sample size, an alpha value of .05, and an f2 effect size calculated
using the R2 value provided in the analysis results.
To control for the false discovery rate associated with multiple comparisons the Benjamini-Hochberg procedure (Benjamini & Hochberg, 1995) was used to adjust the P values, using a false discovery rate of .1. This rate was chosen to balance the high rate advised in the use of hypotheses that are quite explorative in development (due to a limited evidence base), whilst attempting to ensure potential false positive results are controlled for as stringently as possible.
Associations between verbal fluency and demographic factors
Single linear regressions were calculated to predict phonemic fluency (FAS scores) based on education and gender. Phonemic fluency and
education were significantly correlated (F(1,115) 7.639 p=.007), with an R2
of .066 for the effects of education, an f2 effect size of 0.071 and a power (1-
β) of 0.782. Participants’ predicted FAS score was equal to 6.906 + 1.966 (education) where education is coded as 0 = no exams, 1 = educated. Participants’ mean FAS score increased by .091 for participants who were educated. Phonemic fluency and gender were not significantly correlated (F(1,115) 1.664 p=.200), with an R2 of .014 for the effects of gender, an f2
effect size of 0.014 and a power (1-β) of 0.248. Participants’ predicted FAS score was equal to 7.880 + .821 (gender) where gender is coded as 0 =
males, 1 = females. Participants’ mean FAS score increased by .821 for participants who were female.
A single linear regression was calculated to predict semantic fluency (ANT scores) based on education. The regression approached significance (F(1,108) 3.422 p=.067), with an R2 of .031 for the effects of education, an f2
effect size of 0.032 and a power (1-β) of 0.453. Participants’ predicted ANT score was equal to 7.500 + 1.410 (education) where education is coded as 0 = no exams, 1 = educated. Participants’ mean ANT score increased by 1.410 for participants who were educated.
Associations between verbal fluency and cognitive factors Single linear regressions were calculated to predict phonemic fluency (FAS scores) and semantic fluency (ANT) based on semantic memory (GNT scores). A significant regression equation was found for phonemic fluency and semantic memory (F(1,115)13.702, p=.000), with an R2 of .106, an f2
effect size of 0.119 and a power (1-β) of 0.956. Participants’ predicted FAS score was equal to 3.996 + 0.223 (GNT) points. Participants’ FAS score increased by .223 for each point of increase on the GNT. A significant regression equation was also found for semantic fluency and semantic memory (F(1,115) 20.523, p=.000), with an R2 of .151, an f2 effect size of
0.178 and a power (1-β) of 0.994. Participants’ predicted ANT score was equal to 3.001 + .282 (GNT) points. Participants’ ANT score increased by .282 for each point of increase on the GNT.
A single linear regression was calculated to predict phonemic fluency (FAS scores) based on premorbid functioning (TOPF scores). Phonemic fluency and premorbid functioning were significantly correlated
(F(1,108)14.106, p=.000), with an R2 of .109 for the effects of premorbid
functioning, an f2 effect size of 0.122 and a power (1-β) of 0.961.
Participants’ predicted FAS score was equal to 4.717 + 0.091 (TOPF) points. Participants’ FAS score increased by .091 for each point of increase on the TOPF.
Associations between verbal fluency and tumour factors A single linear regression was calculated to predict phonemic fluency (FAS scores) based on tumour type. Phonemic fluency and tumour type were not significantly correlated (F(1,102) .428, p=.514), with an R2 of .004
for the effects of tumour type, an f2 effect size of 0.004 and a power (1-β) of
0.097. Participants’ predicted FAS score was equal to 7.880 + .821 (tumour type) where tumour type was coded as 1 = meningioma, 2 = glioma.
Participants’ mean FAS score increased by .545 for participants with
meningioma in comparison to those with glioma. There were no hypotheses based on tumour factors and semantic fluency.
Associations between verbal fluency and mood factors Single linear regressions were calculated to predict phonemic fluency (FAS scores) and semantic fluency (ANT scores) based on depression (HADS-D) and anxiety (HADS-A). The regression outcomes for phonemic fluency and depression approached significance (F(1,115) 3.857, p=.052), with an R2 of .032 for the effects of depression, an f2 effect size of 0.033 and
a power (1-β) of 0.489. Participants’ predicted FAS score was equal to 8.792 - 1.292 (depression) where depression was coded as 0 = not depressed, 1 = depressed. Participants’ mean FAS score decreased by 1.292 for
participants who were coded as depressed in comparison to those who were not. Phonemic fluency and anxiety were not significantly correlated (F(1,115) .019, p=.890), with an R2 of .000 for the effects of anxiety, an f2 effect size of
0.033 and a power (1-β) of 0.489. Participants’ predicted FAS score was equal to 8.302 + .089 (anxiety) where anxiety was coded as 0 = ‘not
anxious’, 1 = ‘anxious’. Participants’ mean FAS score increased by .089 for participants who were anxious in comparison to those who were not.
The single linear regression calculated to predict semantic fluency (ANT) based on depression (HADS-D) approached significance (F(1,115) 3.249, p=.074), with an R2 of .027 for the effects of depression, an f2 effect
size of 0.370 and a power (1-β) of 0.999. Participants’ predicted ANT score was equal to 8.935 - 1.260 (depression) where depression was coded as 0 = not depressed, 1 = depressed. Participants’ mean ANT score decreased by 1.260 for participants who were coded as depressed in comparison to those
who were not. Semantic fluency and anxiety were not significantly correlated (F(1,115) .461, p=.498), with an R2 of .004 for the effects of anxiety, an f2
effect size of 0.004 and a power (1-β) of 0.103. Participants’ predicted ANT score was equal to 8.755 - .458 (anxiety) where anxiety was coded as 0 = ‘not anxious’, 1 = ‘anxious’. Participants’ mean ANT score increased by -.458 for participants who were anxious in comparison to those who were not.
Associations between verbal fluency and localisation A multiple regression was calculated to predict phonemic fluency based on localisation, with Left Frontal being the reference condition. A significant regression equation was found (F(5,85)3.552, p=.006), with an R2
of .173, an f2 effect size of 0.209 and a power (1-β) of 0.924 for the effects of
localisation. The predicted FAS score for participants with tumours situated in the left frontal lobe was equal to 7.568 (CI 6.575-8.526) given that the other localisation areas take the value of zero. Predicted FAS score
increased by an additional 2.307 points, on average, if the tumour was in the right frontal lobe, or 3.098 points if the tumour was in the right parietal lobe compared with the results for the left frontal lobe. These relationships were shown to be significant (p = .007 and p = .035 respectively). A similar relationship was seen in participants who have tumours in the frontal area, showing an increase of 1.932, but this was not significantly different to the results for left frontal lobe (p = .184). Predicted FAS score decreased by an additional 2.318 points if the tumour was in the left parietal lobe, and 1.425 points if the tumour was in the left temporal lobe. These relationships were not significant (p = .184 and p = .294 respectively).
Table 6. Predictor coefficients for linear regression analyses predicting FAS scores for participants with tumours in different localisations
Unstandardized Coefficients Standardized Coefficients t Sig. B Std. Error β (Constant) 7.568 .500 15.146 .000 Frontal 1.932 1.442 .136 1.339 .184 Left Parietal -2.318 1.731 -.135 -1.339 .184 Left temporal -1.425 1.349 -.108 -1.057 .294 Right Frontal 2.307 .841 .289 2.743 .007 Right Parietal 3.098 1.442 .218 2.148 .035
A multiple regression was calculated to predict semantic fluency based on localisation, with Left Frontal being the reference condition. The results highlight that the regression equation was not significant (F(5,79) .932, p=.465), with an R2 of .056, an f2 effect size of 0.059 and a power (1-β)
of 0.342 for the effects of localisation. The coefficients are summarised in table 7.
Table 7. Predictor coefficients for linear regression analyses predicting ANT scores for participants with tumours in different localisations
Unstandardized Coefficients Standardized Coefficients t Sig. B Std. Error β (Constant) 8.167 .582 14.040 .000 Frontal 1.833 1.783 .115 1.028 .307 Left Parietal .083 1.973 .005 .042 .966 Left Temporal -.738 1.539 -.054 -.480 .633 Right Frontal 1.786 1.007 .206 1.772 .080 Right Parietal .167 1.645 .011 .101 .920
False discovery rate control
Due to the multiple comparisons conducted throughout the individual
variable linear regression analyses a correction was required to account for the potential false discovery rate. The Benjamini-Hochberg Procedure
(Benjamini & Hochberg, 1995) was applied to the P values calculated for the analyses conducted with the FAS (see table 8) and ANT (see table 9).
Table 8. Benjamini-Hochberg outcomes for FAS analyses Variable correlated with FAS Original P- values Benjamini- Hochberg P- values Benjamini-Hochberg significance (0.1 error rate) TOPF 0 0 significant GNT 0 0 significant
Left Frontal 0 0 significant
Education 0.007 0.018 significant
Right Frontal 0.007 0.018 significant Right Parietal 0.035 0.076 significant
HADS-D 0.052 0.097 significant
Frontal 0.184 0.26 not significant
Left Parietal 0.184 0.26 not significant
Gender 0.2 0.26 not significant
Left Temporal 0.294 0.347 not significant Tumour type 0.514 0.557 not significant
HADS-A 0.89 0.89 not significant
After controlling for the error rate using the Benjamini-Hochberg Procedure (Benjamini & Hochberg, 1995) on the analyses conducted for the FAS the significance of the HADS-D correlation changed from not significant (at an α level of .05) to significant (at an error rate of 0.1). While most of the other variables had altered values, they all remained significant or non-significant in line with the significance outcome of the original value.
Table 9. Benjamini-Hochberg outcomes for ANT analyses Variable correlated with ANT Original P- values Benjamini- Hochberg P- values Benjamini-Hochberg significance (0.1 error rate) GNT 0 0 significant
Left Frontal 0 0 significant
Education 0.067 0.16 not significant
HADS-D 0.074 0.16 not significant
Right Fontal 0.08 0.16 not significant
Frontal 0.307 0.512 not significant
HADS-A 0.498 0.711 not significant
Left Temporal 0.633 0.791 not significant Right Parietal 0.92 0.966 not significant Left Parietal 0.966 0.966 not significant
After controlling for the error rate using the Benjamini-Hochberg Procedure (Benjamini & Hochberg, 1995) on the analyses conducted for the ANT all variables remained significant or non-significant in line with the significance outcome of the original value.
Further analyses with phonemic fluency
Hierarchical regression analyses were performed to assess the value of associations between the predictor variables gender, education, tumour type, depression, semantic memory, and premorbid functioning on the relationship between localisation and phonemic fluency. Localisation (with Left Fontal as the reference category) was entered into the first stage of the analysis (block one) to allow for comparison of the effects of the other variables during each stage of the model. Localisation predicted 17.2% of the variance in phonemic fluency scores. The predictor variables (gender, education, tumour type, depression, semantic memory, and premorbid functioning) were added to the second stage of the analysis (block 2). Following the additions in block 2 the total variance explained by the model was 47%, F (6, 68) 6.394, p = 0.000. This means that the predictor variables (gender, education, tumour type, depression, semantic memory, and
premorbid functioning) explained an additional 29.9% of the variance in phonemic fluency scores after localisation had been accounted for. In model 2 there were statistically significant relationships between FAS and a
number of the predictors added after localisation. Gender (β = .228, p = .030) and education (β =.192, p = .047) were statistically significant demographic predictors. Depression (β = -.200, p = .034) and GNT (β = .274, p = .012) were statistically significant psychometric/neuropsychological assessment predictors. The TOPF score variable approached significance (β = .197, p = .077). Table 10 summarises the predictor coefficients for this analysis.
Table 10. Predictor coefficients for hierarchical multiple regression analyses predicting FAS scores
Model
Unstandardized
Coefficients Standardized Coefficients
B
Std.
Error Beta t Sig. 1 (Constant) 7.718 .543 14.216 .000 Frontal 1.082 1.610 .073 .672 .504 Left Parietal -2.468 1.780 -.150 -1.387 .170 Left Temporal -1.575 1.392 -.124 -1.132 .261 Right Frontal 2.335 .949 .277 2.461 .016 Right Parietal 2.949 1.487 .217 1.983 .051 2 (Constant) -.845 1.917 -.441 .661 Frontal .817 1.381 .055 .592 .556 Left Parietal -1.067 1.559 -.065 -.684 .496 Left Temporal .267 1.308 .021 .204 .839 Right Frontal 2.123 .803 .252 2.642 .010 Right Parietal 3.721 1.294 .274 2.877 .005 TOPF .055 .031 .197 1.796 .077 GNT .192 .074 .274 2.576 .012 Depressed -1.473 .679 -.200 -2.169 .034 Gender 1.636 .740 .228 2.210 .030 Education 1.518 .749 .192 2.027 .047 Tumour type .870 .909 .099 .957 .342
Post hoc power analysis of phonemic fluency hierarchical regression
G*Power (Faul et al., 2009) was used to calculate a post hoc analysis for a multiple regression with eleven predictors, an alpha value of .05, a sample size of 80 and an f2effect size of 0.887 as calculated using the R 2
the power achieved in this analysis was 0.99, which is a very high power score, indicating that the likelihood of the results being due to a type 2 error are very low.
Further analyses with semantic fluency
Hierarchical regression analyses were performed to assess the value of associations between the predictor variables education, depression, and semantic memory on the relationship between localisation (with Left Frontal as the reference condition) and semantic fluency. Localisation was entered into the first stage of the analysis (block one) to allow for comparison of the effects of the other variables during each stage of the model. Localisation did not predict variance in semantic fluency scores (F(5,79) .932, p=.465), with an R2 of .056. The predictor variables (education, depression, and
semantic memory) were added to the second stage of the analysis (block 2). Following the additions in block 2 the total variance explained by the model was 24%, F (3, 76) 6.158, p = 0.001. This means that the predictor variables (education, depression, and semantic memory) explained an additional 18.5% of the variance in semantic fluency scores after localisation had been accounted for. In model 2 there was a statistically significant relationship between ANT and GNT (β = .360, p = .001). The association between ANT and depression approached significance (β = -.176, p = .086). Table 11 summarises the predictor coefficients for this analysis.
Table 11. Predictor coefficients for hierarchical multiple regression analyses predicting ANT scores
Model
Unstandardized
Coefficients Standardized Coefficients
B
Std.
Error Beta t Sig. 1 (Constant) 8.167 .582 14.040 .000 Frontal 1.833 1.783 .115 1.028 .307 Left Parietal .083 1.973 .005 .042 .966 Left Temporal -.738 1.539 -.054 -.480 .633 Right Frontal 1.786 1.007 .206 1.772 .080 Right Parietal .167 1.645 .011 .101 .920 2 (Constant) 2.414 1.767 1.366 .176 Frontal .912 1.649 .057 .553 .582 Left Parietal 1.059 1.836 .060 .577 .566 Left Temporal .828 1.465 .061 .565 .574 Right Frontal 1.475 .928 .170 1.589 .116 Right Parietal .312 1.513 .021 .207 .837 GNT .269 .079 .360 3.397 .001 Depressed -1.366 .784 -.176 -1.742 .086 Education 1.349 .873 .160 1.545 .126
Post hoc power analysis of semantic fluency hierarchical regression
G*Power (Faul et al., 2009) was used to calculate a post hoc analysis for a multiple regression with eleven predictors, an alpha value of .05, a sample size of 85 and an f2effect size of 0.316 as calculated using the R 2
value provided in the analysis results (0.240). This calculation indicated that the power achieved in this analysis was 0.994, which is a very high power score, indicating that the likelihood of the results being due to a type 2 error is very low.
Summary of results
An increase in phonemic fluency was significantly correlated with being educated (F(1,115) 7.639 p=.007; R2 = .066), an increase in semantic
memory (F(1,115)13.702, p=.000; R2 = .106), and an increase in premorbid
functioning (F(1,108)14.106, p=.000; R2 = .109). Phonemic fluency was
significantly correlated with localisation (F(5,85)3.552, p=.006; R2 = .173).
More specifically, an increase in phonemic fluency was significantly
associated with tumours in the right frontal and right parietal lobes (p = .007 and p = .035 respectively) in comparison to the left frontal lobe. Phonemic fluency was not significantly correlated with depression (F(1,115) 3.857, p=.052; R2 = .032) in the original regression analysis, however when the
Benjamini-Hochberg Procedure was applied to account for the false
discovery rate phonemic fluency was significantly correlated with depression (0.097; using an error rate of 0.1).
An increase in semantic fluency was correlated with an increase in semantic memory (F(1,115) 20.523, p=.000; R2 = .151).
Phonemic fluency was not significantly correlated with gender
(F(1,115) 1.664 p=.200; R2 = .014), tumour type (F(1,102) .428, p=.514; R2 =
.004), or anxiety (F(1,115) .019, p=.890; R2 = .000). Neither was phonemic
fluency significantly correlated with tumours situated in the frontal lobe (p = .184), left parietal lobe (p = .184), or the left temporal lobe (p = .294) in comparison to the left frontal lobe.
Semantic fluency was not correlated with education (F(1,108) 3.422 p=.067; R2 = .031), or depression (F(1,115) 3.249, p=.074; R2 = .027),
however, there were trends indicating that an increase in semantic fluency was associated with being educated and a decrease in depression scores. Semantic fluency was not correlated with anxiety (F(1,115) .461, p=.498; R2
= .004) or localisation (F(5,79) .932, p=.465; R2 = .056).
Discussion
This chapter summarises the above results and discusses the hypotheses tested, evaluating the strengths and limitations of the study in relation to the current literature. The clinical utility of the results will be
considered and suggestions for future studies provided. This section will begin with a summary of the findings and then discuss the findings for each area of interest (demographic factors, cognitive factors, tumour factors, mood factors and localisation) in more detail. Following the discussion about the outcomes this section will go on to consider the strengths and
weaknesses of this study, concluding with reflections on the clinical implications of these outcomes and suggestions for future research.